Vulnerabilities and Protections in Large Language Models

Abstract

Large Language Model (LLM) security includes monitoring an LLM; detecting an attack on the LLM and defining an attack type of a plurality of attack types based on the monitoring, providing a notification of the attack; and causing a defense to the attack based on the attack type. Advantageously, the security can be configured to be executed between a user outside of the LLM. Further, the security can be configured to defend against multi-turn attacks.

What is claimed is: 
     
         1 . A method for large language model security comprising steps of:
 inline monitoring a Large Language Model (LLM);   detecting an attack on the LLM and defining an attack type of a plurality of attack types based on the monitoring;   providing a notification of the attack; and   causing a defense to the attack based on the attack type.   
     
     
         2 . The method of  claim 1 , wherein the monitoring includes monitoring a user input to the LLM. 
     
     
         3 . The method of  claim 1 , wherein the defense includes blocking the user input to the LLM. 
     
     
         4 . The method of  claim 1 , wherein the plurality of attack types includes prompt hacking and adversarial attack. 
     
     
         5 . The method of  claim 4 , wherein prompt hacking is one of a prompt injection and a jailbreaking attack. 
     
     
         6 . The method of  claim 4 , wherein the adversarial attack is one of a backdoor attack and a data poisoning attack. 
     
     
         7 . The method of  claim 1 , wherein causing the defense includes causing any of a prevention-based defense and a detection-based defense. 
     
     
         8 . The method of  claim 1 , wherein any of the monitoring, detecting, providing a notification, and causing the defense is performed by an intermediate system before a query reaches the large language model. 
     
     
         9 . The method of  claim 1 , wherein the defense includes one of removing, altering, and redesigning an output. 
     
     
         10 . The method of  claim 1 , wherein the detection includes one of response-based detection and prompt-based detection. 
     
     
         11 . The method of  claim 1 , wherein the defense includes one of system-mode self-reminder prompts, smooth LLM, black-box defense, and pretrained language model defense. 
     
     
         12 . A non-transitory computer-readable medium comprising instructions that, when executed, cause one or more processors to perform steps of:
 inline monitoring a large language model (LLM);   detecting an attack defining an attack type of a plurality of attack types on the large language model based on the monitoring;   providing a notification of the attack; and   causing a defense to the attack based on a type of the attack type.   
     
     
         13 . The non-transitory computer-readable medium of  claim 12 , wherein the defense includes blocking a user input to the LLM. 
     
     
         14 . The non-transitory computer-readable medium of  claim 12 , wherein the plurality of attack types includes prompt hacking and adversarial attack. 
     
     
         15 . The non-transitory computer-readable medium of  claim 12 , wherein causing the defense comprising causing any of prevention-based defense and detection-based defense. 
     
     
         16 . The non-transitory computer-readable medium of  claim 12 , wherein any of the monitoring, the detecting, providing a notification, and causing the defense is performed by an intermediate system before a query reaches the large language model. 
     
     
         17 . The non-transitory computer-readable medium of  claim 12 , wherein the defense includes one of removing, altering, and redesigning an output. 
     
     
         18 . The non-transitory computer-readable medium of  claim 12 , wherein the detection includes one of response-based detection and prompt-based detection. 
     
     
         19 . The non-transitory computer-readable medium of  claim 12 , wherein the defense includes one of system-mode self-reminder prompts, smooth LLM, black-box defense, and pretrained language model defense. 
     
     
         20 . The non-transitory computer-readable medium of  claim 12 , wherein the defense includes a three step clustering based defense comprising representation learning, clustering, and filtering.